fruffy/flay

flay's Issues

Shall we allow adding disjunctive condition to reachabilityMap in addReachabilityMapping?

In current implementation of addReachabilityMapping, we always try to initialize for a pair of node and cond. However, in cases like this:

action copy_to_cpu(...) {
    // An if statement
    if (some_cond) {
        x = a;
    } else {
        x = b;
    }
    // codes to copy to cpu
}

action redirect_to_cpu(...) {
    // codes to drop
    copy_to_cpu(...) // this calls the previous action
}

the if statement in copy_to_cpu will be visited twice. And our ReachabilityMap is compared by SourceIdCmp. Thus, the second time to call addReachabilityMapping for this if statement will failed here.

I encountered this issue when trying on Tofino's switch program. And my fix in mind is instead of asserting existence of node, we should add new condition as disjunction to the reachabilityMap.

Add reference files.

We should keep track off what kind of code we consider dead.

Implement a data plane variable and collapse any operation involving this variable.

Use taint tracking for this. We consider data plane variables arbitrary so maybe we can get some performance by collapsing them.

Does current `table_executor` consider default action reachability?

The current implementation of table_executor adds reachability for each action here. But the expression actionHitCondition is constructed with new IR::LAnd(hitCondition, actionChoice), where hitCondition is for key matching and actionChoice is for action name.

But for default actions, shall we add something like "if none of the actions matched, the default action will be reached"?

The reason I am thinking of this is that I noticed the removing for tna_operations.p4 looks incorrect. For example, if a packet misses on table forward, it should be dropped by executing default action miss. But since forward.apply() is removed, it won't be dropped.

Flay does not model default action override.

It is possible to override a non-constant default action and create a new path. We do not model this behavior yet.

We do not need a separate implementation for initial entries in the data plane.

We can just generate the necessary entries from the control-plane and reject any other entries.